Matched two-case free-time test

Fixed question and source evidence

Stretch 0e73879f98c3b701fd7c42df has two competing saved alignments in each processing case. The completed matched fixed-time study reproduced all eight restricted fits and passed all four within-anchor repeat comparisons. Timing gradients remain nonzero at some endpoints, so restricted completion alone cannot resolve the stretch.

This study uses those sealed results, not the interrupted free-time study of a different stretch. All prior failures and receipts remain unchanged.

Implementation

analysis/tests/ilvis0_remaining_free_time_v2_worker.py privately adapts the completed profile-seeded executor. Its original source is never edited. The existing lll.ilvis0_profile_seeded candidate and bounded trust solver are reused. New tests check exact case/start coverage, missing or failed profiles, mixed layouts, unchanged modules, separate repeat groups and accounting. Existing known-answer controls exercise free timing, bounds, failure propagation and the real immutable receipt ledger.

Before launch, seal the successful software qualification, plan, source snapshots, original-file hashes, saved evidence, native kernel and numerical environment. Hold exclusive locks on the worker and source study. Every native call is charged before execution; an incomplete receipt transaction blocks continuation. Charged starts cannot be retried, and no completed allowance is reused.

Finite allowance

The 81 fitting calls include one reproduction of the selected saved endpoint and at most 80 solver calls. At approximately 22 seconds per native call, exhausting the allowance takes about two hours. Earlier stopping is possible.

Unchanged physical assumptions

Both cases retain ±1 degree/hour constant gyro offsets and either fixed-zero or profiled upstream background-field removal. The nominal 200 Hz increment clock, receiver coordinate weights, selected interval, ±0.15-second alignment bounds, other physical bounds and zero-field globe model are unchanged. No raw measurement is interpolated or repaired.

For each case and starting solution, retain both verified profile anchors and choose the smaller saved objective, with alignment as a deterministic tiebreak. No result from the other case enters seed selection. Release all original parameters; do not hold timing fixed during the new optimization.

Required checks and interpretation

Both fresh controls must pass before target starts. Every target must pass the existing full stationarity, repeated-center and fresh physical-position stability checks. Reproduce the selected saved objective and final objective without an unexplained increase. Save final parameters, residuals, full Jacobian and packet boundaries for every completed check.

Compare primary and repeat separately within each processing case using the unchanged relative-cost limit of 10⁻⁶ and 0.01-metre per-axis track RMS. Report failures as failures, including supported integration-boundary problems. Do not replace smooth full-space checks with relaxed tolerances.

Even a successful trial requires a separate saved companion/model review before changing a published disposition. All six primary model fits, fresh controls and required repeat agreements remain necessary. No global optimality, calibrated significance, processing provenance or causal interpretation follows from this numerical test. The other two unresolved stretches are outside its allowance.

Shared source directories acquire one exclusive lock each. The earlier preparation remains preserved with zero charged starts and zero native calls; its duplicate-lock startup failure cannot be resumed. This separately frozen launcher changes lock handling only and retains the same finite scientific scope.